GridGEN: generative models for distribution grids

Master’s thesis, Professorship of Energy Management Technologies, TUM · supervised by Ehimare Okoyomon · grade 1.0 · published at the NeurIPS 2025 workshop Tackling Climate Change with Machine Learning

Problem

Machine learning for power distribution networks is held back by a lack of public grid data. Synthetic grids that are structurally realistic would make it easier to benchmark and share methods.

Approach

• A variational graph autoencoder (PyTorch Geometric) that learns a latent representation of grid topologies and samples new ones.
• Four decoders compared: inner product, MLP, GCN, and a self-designed Iterative-GCN with a refinement loop.
• Evaluated on two open datasets: ENGAGE (3,000 training grids, ≈100 nodes) and DINGO (2,722 training grids, 4,500–7,000 nodes), using average degree and the Wasserstein distance between Laplacian spectra.

Findings

• Simple decoders fail to capture realistic topologies.
• GCN-based decoders reach strong fidelity on ENGAGE. Iterative-GCN gave the best structural fidelity on both datasets, and on ENGAGE matched the real average degree within 1% (2.07 vs. 2.05).
• The larger DINGO grids remain hard: generated networks show disconnected components and repeated motifs. This points to the need for more expressive generative models and stronger evaluation.

Models and analysis are released as open source.